Apertus 1.5 Release Notes

Apertus 1.5, the latest release of the Apertus models, introduces several enhancements including image and audio understanding, an optional thinking mode, and a significantly longer context window. These updates were made possible by pretraining the models on 4 trillion text tokens and 2 trillion multimodal tokens, resulting in updated 8B and 70B model weights.

The new models use an early-fusion multimodality approach, where images and audio are tokenized alongside text tokens in the same sequence. This allows the models to understand images and audio, while still generating only text output. The Apertus 1.5 models are fully open source, with open weights, data, and training details. Post-training evaluations have been expanded to include multimodal and thinking-mode evaluations, demonstrating strong performance on coding, instruction following, and reasoning tasks.

Apertus 1.5 also shows competitive performance in spatial understanding and document/chart understanding among open-weight models. The release continues Apertus' tradition of transparency, with all evaluation results published in the Eval Card format, which includes details about the evaluation process and model access. This commitment to transparency aligns with AI governance frameworks such as the EU AI Act and the GPAI Code of Practice.

All data and training details for Apertus 1.5 are publicly available, allowing researchers to investigate the model's behavior and compare it to other models.